Skip to content

Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.

fix(settings): fall back to /v1 when listing self-hosted models (#1235)

Gabriel Steinbee0f67632c4
diff --git a/src/components/OpenAICompatiblePanel.tsx b/src/components/OpenAICompatiblePanel.tsx
index ab89b7a57..6fdb74567 100644
--- a/src/components/OpenAICompatiblePanel.tsx
+++ b/src/components/OpenAICompatiblePanel.tsx
@@ -5,7 +5,7 @@ import { Input } from "./ui/input";
import ApiKeyInput from "./ui/ApiKeyInput";
import ModelCardList from "./ui/ModelCardList";
import SearchableModelList, { MODEL_SEARCH_THRESHOLD } from "./ui/SearchableModelList";
-import { buildApiUrl, normalizeBaseUrl } from "../config/constants";
+import { buildApiUrl, getModelListBaseCandidates, normalizeBaseUrl } from "../config/constants";
import { isSecureEndpoint } from "../utils/urlUtils";
import { GetApiKeyLink } from "./ui/GetApiKeyLink";
@@ -111,6 +111,7 @@ export default function OpenAICompatiblePanel({
const trimmedKey = apiKey?.trim();
const effectiveKey = trimmedKey && trimmedKey.length > 0 ? trimmedKey : undefined;
+ let activeBase = normalized;
try {
if (!normalized.includes("://")) {
@@ -134,50 +135,79 @@ export default function OpenAICompatiblePanel({
headers.Authorization = `Bearer ${effectiveKey}`;
}
- const modelsUrl = buildApiUrl(normalized, "/models");
- const response = await fetch(modelsUrl, { method: "GET", headers });
+ const fetchModelOptions = async (base: string): Promise<ModelOption[]> => {
+ const response = await fetch(buildApiUrl(base, "/models"), { method: "GET", headers });
- if (!response.ok) {
- const errorText = await response.text().catch(() => "");
- const summary = errorText
- ? `${response.status} ${errorText.slice(0, 200)}`
- : `${response.status} ${response.statusText}`;
- throw new Error(summary.trim());
+ if (!response.ok) {
+ const errorText = await response.text().catch(() => "");
+ const summary = errorText
+ ? `${response.status} ${errorText.slice(0, 200)}`
+ : `${response.status} ${response.statusText}`;
+ throw new Error(summary.trim());
+ }
+
+ const payload = await response.json().catch(() => ({}));
+ const rawModels = Array.isArray(payload?.data)
+ ? payload.data
+ : Array.isArray(payload?.models)
+ ? payload.models
+ : [];
+
+ // Coerce fields defensively: non-conformant endpoints may return
+ // numeric ids or object descriptions, which would crash the render.
+ return (rawModels as Array<Record<string, unknown>>)
+ .map((item) => {
+ const rawValue = item?.id ?? item?.name;
+ if (rawValue === undefined || rawValue === null || rawValue === "") return null;
+ const value = String(rawValue);
+ const ownedBy = typeof item?.owned_by === "string" ? item.owned_by : undefined;
+ const description =
+ typeof item?.description === "string" && item.description
+ ? item.description
+ : ownedBy
+ ? t("reasoning.custom.ownerLabel", { owner: ownedBy })
+ : undefined;
+ return { value, label: value, description, ownedBy } as ModelOption;
+ })
+ .filter(Boolean) as ModelOption[];
+ };
+
+ // When the entered base yields no models, fall back through sibling
+ // bases and adopt the working one so inference targets it too.
+ const candidates = lockedBaseUrl ? [normalized] : getModelListBaseCandidates(normalized);
+ let mapped: ModelOption[] = [];
+ let resolvedBase = normalized;
+ let primaryError: Error | null = null;
+
+ for (const candidate of candidates) {
+ try {
+ const options = await fetchModelOptions(candidate);
+ if (options.length > 0) {
+ mapped = options;
+ resolvedBase = candidate;
+ break;
+ }
+ } catch (error) {
+ if (candidate === normalized) primaryError = error as Error;
+ }
}
- const payload = await response.json().catch(() => ({}));
- const rawModels = Array.isArray(payload?.data)
- ? payload.data
- : Array.isArray(payload?.models)
- ? payload.models
- : [];
-
- // Coerce fields defensively: non-conformant endpoints may return
- // numeric ids or object descriptions, which would crash the render.
- const mapped = (rawModels as Array<Record<string, unknown>>)
- .map((item) => {
- const rawValue = item?.id ?? item?.name;
- if (rawValue === undefined || rawValue === null || rawValue === "") return null;
- const value = String(rawValue);
- const ownedBy = typeof item?.owned_by === "string" ? item.owned_by : undefined;
- const description =
- typeof item?.description === "string" && item.description
- ? item.description
- : ownedBy
- ? t("reasoning.custom.ownerLabel", { owner: ownedBy })
- : undefined;
- return { value, label: value, description, ownedBy } as ModelOption;
- })
- .filter(Boolean) as ModelOption[];
+ if (mapped.length === 0 && primaryError) throw primaryError;
if (isMountedRef.current && latestBaseRef.current === normalized) {
+ if (resolvedBase !== normalized) {
+ activeBase = resolvedBase;
+ latestBaseRef.current = resolvedBase;
+ setDraftBase(resolvedBase);
+ setBaseUrl(resolvedBase);
+ }
setModelOptions(mapped);
// `/models` is a discovery aid, not an allowlist — keep the user's
// chosen id even if it's absent (it may still be valid, or belong to
// a provider they're switching away from). Invalid ids surface a
// clear API error at request time instead of being silently wiped.
setModelsError(null);
- lastLoadedBaseRef.current = normalized;
+ lastLoadedBaseRef.current = resolvedBase;
}
} catch (error) {
if (isMountedRef.current && latestBaseRef.current === normalized) {
@@ -194,12 +224,12 @@ export default function OpenAICompatiblePanel({
if (pendingBaseRef.current === normalized) {
pendingBaseRef.current = null;
}
- if (isMountedRef.current && latestBaseRef.current === normalized) {
+ if (isMountedRef.current && latestBaseRef.current === activeBase) {
setModelsLoading(false);
}
}
},
- [baseUrl, apiKey, t]
+ [baseUrl, apiKey, lockedBaseUrl, setBaseUrl, t]
);
useEffect(() => {
diff --git a/src/config/constants.ts b/src/config/constants.ts
index 9ce6cb4bf..fd4b093fc 100644
--- a/src/config/constants.ts
+++ b/src/config/constants.ts
@@ -43,6 +43,19 @@ export const ensureV1Suffix = (base: string): string => {
return normalized.endsWith("/v1") ? normalized : `${normalized}/v1`;
};
+// Ordered bases to try when listing models from an OpenAI-compatible server.
+// Self-hosted servers (LM Studio, Ollama, vLLM) serve the API under /v1 even
+// when users enter the bare origin, and LM Studio's native REST base
+// (/api/v1 or /api/v0) has its OpenAI-compatible sibling at /v1.
+export const getModelListBaseCandidates = (base: string): string[] => {
+ const normalized = normalizeBaseUrl(base);
+ if (!normalized) return [];
+ const nativeApiMatch = normalized.match(/^(.+?)\/api\/v[01]$/i);
+ if (nativeApiMatch) return [normalized, `${nativeApiMatch[1]}/v1`];
+ const withV1 = ensureV1Suffix(normalized);
+ return withV1 === normalized ? [normalized] : [normalized, withV1];
+};
+
const env = (typeof import.meta !== "undefined" && (import.meta as any).env) || {};
const computeBaseUrl = (candidates: Array<string | undefined>, fallback: string): string => {
diff --git a/test/helpers/modelListBaseCandidates.test.js b/test/helpers/modelListBaseCandidates.test.js
new file mode 100644
index 000000000..bb389c488
--- /dev/null
+++ b/test/helpers/modelListBaseCandidates.test.js
@@ -0,0 +1,62 @@
+const test = require("node:test");
+const assert = require("node:assert/strict");
+
+// Requires Node's native TypeScript type-stripping (Node >= 22.6 with
+// --experimental-strip-types, on by default in Node 23.6+/24). CI runs Node 24.
+
+const load = () => import("../../src/config/constants.ts");
+
+test("bare origin falls back to /v1 (LM Studio, Ollama, vLLM)", async () => {
+ const { getModelListBaseCandidates } = await load();
+
+ assert.deepEqual(getModelListBaseCandidates("http://127.0.0.1:1234"), [
+ "http://127.0.0.1:1234",
+ "http://127.0.0.1:1234/v1",
+ ]);
+});
+
+test("LM Studio native REST bases fall back to the OpenAI-compatible /v1", async () => {
+ const { getModelListBaseCandidates } = await load();
+
+ assert.deepEqual(getModelListBaseCandidates("http://127.0.0.1:1234/api/v1"), [
+ "http://127.0.0.1:1234/api/v1",
+ "http://127.0.0.1:1234/v1",
+ ]);
+ assert.deepEqual(getModelListBaseCandidates("http://127.0.0.1:1234/api/v0"), [
+ "http://127.0.0.1:1234/api/v0",
+ "http://127.0.0.1:1234/v1",
+ ]);
+});
+
+test("bases already ending in /v1 have no fallback", async () => {
+ const { getModelListBaseCandidates } = await load();
+
+ assert.deepEqual(getModelListBaseCandidates("http://127.0.0.1:1234/v1"), [
+ "http://127.0.0.1:1234/v1",
+ ]);
+ assert.deepEqual(getModelListBaseCandidates("https://api.together.xyz/v1"), [
+ "https://api.together.xyz/v1",
+ ]);
+});
+
+test("hosted /api/v1 bases keep the entered base first so it wins when it serves models", async () => {
+ const { getModelListBaseCandidates } = await load();
+
+ assert.deepEqual(getModelListBaseCandidates("https://openrouter.ai/api/v1"), [
+ "https://openrouter.ai/api/v1",
+ "https://openrouter.ai/v1",
+ ]);
+});
+
+test("normalizes trailing slashes and endpoint suffixes before deriving candidates", async () => {
+ const { getModelListBaseCandidates } = await load();
+
+ assert.deepEqual(getModelListBaseCandidates("http://127.0.0.1:1234/"), [
+ "http://127.0.0.1:1234",
+ "http://127.0.0.1:1234/v1",
+ ]);
+ assert.deepEqual(getModelListBaseCandidates("http://127.0.0.1:1234/v1/models"), [
+ "http://127.0.0.1:1234/v1",
+ ]);
+ assert.deepEqual(getModelListBaseCandidates(""), []);
+});